Acoustic Emission Signatures as Predictors of Laser Weld Quality: A Machine Learning Framework for Real-Time Laser Welding Quality Assessment

Ensuring weld integrity in dissimilar-material joints remains a critical challenge, as current quality control relies heavily on destructive post-process inspection. Here, we assess airborne Acoustic Emission (AE) monitoring for in-process quality evaluation in laser welding of cemented carbide-steel. AE recorded over 11 welding conditions that were described by time- and frequency-domain features, partitioned by unsupervised clustering into three regimes, unstable weld (UW), quality weld (QW) and crack-defective weld (CDW), and classified by four shallow Machine Learning models. Because 10 ms windows from a single weld are not independent observations, performance is reported using grouped cross-validation, in which each window from a held-out weld is excluded from training, at two levels: leave-one-reading-out (LORO) and the stricter leave-one-condition-out (LOCO). Crack-defective welds caused by lateral beam offset were identified in all AE readings under both schemes (window-level AUC 0.987–1.000). Incomplete versus continuous bonding was separated with 0.846 reading-level accuracy under LOCO, and a combined two-stage model reached 0.857 accuracy and 0.883 balanced accuracy (permutation 0.002). Amplitude-domain features increased monotonically with heat input and saturated near 35 J/mm, so the acoustic signature varies gradually with heat input rather than forming a distinct class, whereas beam offset altered the spectral signature. Shear strength and fracture surface analyses corroborated the physical interpretation of the three regimes. The results establish that AE encodes an interpretable and predictive fingerprint of weld integrity and delimits the conditions under which it generalises to unseen welds.